Skip to content

Terminal-Bench 2.0

A benchmark for agentic software engineering tasks executed in real terminal environments. DeepSeek reports it in the agentic section, while BenchLM also mirrors it in coding for models that publish it as a developer-task signal.

39 models

Higher is better.

  1. 1GPT-5.5OpenAI82.0
  2. 2GPT-5.3 CodexOpenAI77.3
  3. 3GPT-5.4OpenAI75.1
  4. 4Qwen3.7 PlusAlibaba70.3
  5. 5Qwen3.7 MaxAlibaba69.7
  6. 6Claude Opus 4.7 (Adaptive)Anthropic69.4
  7. 7Composer 2.5Cursor69.3
  8. 8MiMo-V2.5-ProXiaomi68.4
  9. 9DeepSeek V4 Pro 0813DeepSeek67.9
  10. 10Kimi K2.6Moonshot AI66.7
  11. 11MiMo-V2.5Xiaomi65.8
  12. 12Claude Opus 4.6Anthropic65.4
  13. 13Qwen 3.6 Max (preview)Alibaba65.4
  14. 14GLM-5.1Z.AI63.5
  15. 15Composer 2Cursor61.7
  16. 16Qwen3.6 PlusAlibaba61.6
  17. 17GPT-5.4 miniOpenAI60.0
  18. 18Claude Opus 4.5Anthropic59.3
  19. 19Qwen3.6-27BAlibaba59.3
  20. 20Claude Sonnet 4.6Anthropic59.1
  21. 21Muse SparkMeta59.0
  22. 22MiniMax M2.7MiniMax57.0
  23. 23DeepSeek V4 Flash 0731DeepSeek56.9
  24. 24GLM-5Z.AI56.2
  25. 25Hy3 PreviewTencent54.4
  26. 26Qwen3.5 397BAlibaba52.5
  27. 27Qwen3.6-35B-A3BAlibaba51.5
  28. 28Kimi K2.5 (Reasoning)Moonshot AI50.8
  29. 29Kimi K2.5Moonshot AI50.8
  30. 30Claude Sonnet 4.5Anthropic50.0
  31. 31Qwen3.5-122B-A10BAlibaba49.4
  32. 32Grok 4.20xAI47.1
  33. 33GPT-5.4 nanoOpenAI46.3
  34. 34MAI-Thinking-1Microsoft46.0
  35. 35Laguna M.1Poolside45.8
  36. 36Qwen3.5-27BAlibaba41.6
  37. 37GLM-4.7Z.AI41.0
  38. 38Qwen3.5-35B-A3BAlibaba40.5
  39. 39LongCat-Flash-Lite-SparseMeituan33.7
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Interactive CLI agent evaluation
Tasks
Terminal-based software tasks
Authors
Terminal-Bench contributors (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.